The reproducibility of SELDI spectra, i.e., mass location and intensity from array to array on a single chip (intra-assay) and between chips (inter-assay), was determined using the pooled normal serum quality control samples. Seven protein peaks in the range of 3,000–10,000 m/z on the observed spectra were randomly selected to calculate the coefficient of variance. The intra-assay and inter-assay coefficient of variance for peak location was 0.05%, and the intra-assay and inter-assay coefficients of variance for normalized intensity (peak height or relative concentration) were 6% and 14% respectively (data not shown). Masses that were within 0.08% mass accuracy between spectra were considered to be the same. Most importantly, it was observed that randomly selected samples, blinded to the person performing SELDI and rerun months or even a year later, were correctly classified by the decision tree classification algorithm.

Initially, we analyzed serum samples from the training set, using the Ciphergen Biomarker pattern software. Among 126 qualified mass peaks (signal-to-noise ratio>5; ranging from 2 to 50 km/z), 21 protein peaks were over-expressed, whereas 44 protein peaks were significantly downregulated in sera from HCC patients compared with those from controls (all p values<0.05; Table 1). Fig. 1 shows an example of the 7984 m/z peak that was up-regulated in HCC samples vs. control samples.

To identify biomarkers with the potential to detect HCC, the intensities of protein peaks in training set were transferred to BPS. A total of 6 peaks with the highest discriminatory power (3157.33 m/z, 4177.02 m/z, 4284.79 m/z, 4300.80 m/z, 7789.87 m/z, and 7984.14 m/z) were automatically selected to construct a classification tree (Table 2, 3). Fig. 2 shows the tree structure and sample distribution. This classification model distinguished serum samples based on their profiles of the protein peak intensities. The classification tree using the combination of these 6 protein peaks identified 33 cases of HCC and 33 controls, resulting in a sensitivity of 100% and a specificity of 96.97% respectively. (Table 4)

To evaluate the accuracy and validity of the classification tree generated from the training set, we determined the performance of the classification model in our testing dataset, consisting of 48 HCC and 33 of control samples. Consistent with the results in training set, this classification tree separated HCC samples from control samples with a sensitivity of 100%, a specificity of 96.97%, and a positive predictive value of 95.92%, respectively (Table 4). The area under the receiver operating characteristics (ROC) curve of this model was 0.986, indicating possible diagnostic utility.

To isolate the proteins of interest and to determine candidate protein identities, 6 serum samples from HCC patients containing high SELDI intensity of 7789 M/z and 7984 M/z were selected as potential candidate proteins for isolation by Tricine-SDS-PAGE. Fig. 3 shows a picture of the Tricine-SDS-PAGE gel separating two fractions of proteins with approximate mass between 2.5 km/z and 11 k m/z. The band with 8 m/z mobility that stained more intensively in the gel was excised and trypsinised (Fig. 3).

After trypsin digestion of the 8 k m/z gel band, digested peptides were measured by LC-MS/MS. 79 peptides, including 45 unique peptides, were identified. Fig. 4 and Fig. 5 show the peptide mass fingerprinting of the 7984 m/z band after trypsin digestion. Peptide mass lists derived from spectra with high S/N ratio were submitted to SWISS-PROT/TrAEMBL for database matching. By database search, we found that the sequence of the 7984.14 m/z protein (pI>4) matched with the neutrophil-activating peptide MW = 8 019.36 m/z, pI = 8.43), a protein known for its high expression in liver tumor tissues.

NAP-2 expression in liver tumor tissues and adjacent liver tissues was analyzed by immunohistochemistry using specific NAP-2 antibodies. Interestingly, strong brown staining signals were found in HCC tissues (Fig. 6A), while no NAP-2 staining was found in adjacent liver tissues (Fig. 6B).

With informed consent obtained from each participant, serum samples were collected by the Department of Surgery at the First Affiliated Hospital of Guangxi Medical University, China. Patients with HCC were diagnosed according to standard criteria put forth by the Chinese Society of Liver Cancer 33. The cancer group consisted of 48 patients with HCC. All HCC cases were histologically confirmed, positive for hepatitis B antigen, and negative for anti-hepatitis C. The controls were 33 healthy volunteers without liver neoplasia, alcoholic cirrhosis, hepatitis B or hepatitis C infection, recruited from routine health examination at the same hospital. The mean age was 46.0 years (ranges from 26 to 75 years) for HCC patients and 41.8 years (ranges from 23 to 60 years) for healthy controls. The characteristics of all subjects are shown in Table 5. Sera of HCC patients was collected before any treatment and randomly divided into two groups: training group and testing group. All blood samples were colleted in the morning before breakfast. Two milliliters of whole blood were obtained and stored at 4°C for one hour and then centrifuged for 10 minutes at 3000 r/min. All serum samples were stored at -80°C before SELDI ProteinChip analysis.

Protein profiling of serum samples was performed using the eight-spot format WCX2 (weak cationic exchange) ProteinChip Arrays (Ciphergen Biosystems, Fremont, CA, USA). Frozen serum samples were thawed and spun at 10,000 rpm for 5 min at 4°C. Twenty μl of U9 buffer was added to 10 μl aliquots of each serum sample and placed on ice for 30 min before adding 360 μl WCX-2 buffer. Arrays were prepared as follows: each array was pre-equilibrated 2 × 5 min in 200 μl WCX-2 buffer on a horizontal shaker (MSI Minishaker) before sample addition. The sample supernatant was added and incubated for 1 hr on the shaker. After incubation, the sample was removed, and each spot was washed with 200 μl WCX-2 buffer for 2 × 5 min with agitation. After washing, the array was carefully separated from the bioprocessor and washed briefly with deionized water. 0.5 μl sinapinic acid (SPA) was deposited on the array spots and allowed to air dry.

The ProteinChip Arrays were read by surface-enhanced laser desorption/ionization time-of flight (SELDI-TOF-MS) mass spectrometry (ProteinChip PBS II reader, Ciphergen). This was calibrated using NP20 chips that had been bound with all-in-one standard proteins to set up the parameters. The optimal detection parameter of mass/charge size range was set between 2000 and 10000 M/Z with a maximum of 50000 M/Z. The laser intensity was set at 175 and detector sensitivity was set at 5. An average value of 130 spots was presented for each sample. All samples were detected with the same parameters. All the raw data was normalized with the ProteinChip Software version 3.1 (homogenization of the total ion strength and M/Z). The M/Z sample peaks with more than 2000 M/Z were normalized with biomarker wizard of ProteinChip Software version 3.1 for noise filtering. The first threshold for noise filtering was set at 5, and the second was set at 2. The minimum threshold for clustering was set at10%. Spectrum analysis was performed using the Biomarker Patterns Software.

Patients with HCC were split into a training set and a testing set. The training sample set consisted of 33 HCC patients and 33 healthy controls. The protein profiling spectra obtained from the serum samples were normalized using total ion current normalization from Ciphergen's ProteinChip Software version 3.1. Peak labeling was performed by the Biomarker Wizard feature of this software. The intensities of selected peaks were then transferred to Biomarker Pattern Software (BPS) as a 'root node'. On the basis of peak intensity, a threshold was determined by BPS to classify the root node into two child nodes. If the peak intensity of a blind sample was lower than or equal to the threshold, this peak would be labeled as left-side child node. Peak intensities higher than the threshold would be marked as right-side child node. After multiple rounds of decision-making, BPS pooled all labeled samples into a terminal node, where samples were divided as cancer group or control group using the classification tree. Classification of the training set was made to yield the least classification error.

The testing set consisted of serum samples from 48 patients with HCC and 33 control individuals. Using the classification model generated from training dataset, BPS evaluated all of the protein peak intensities for each sample in the testing dataset. It then discriminated HCC and control samples according to their proteomic profile characteristics. Model sensitivity was defined as the probability of predicting HCC cases, while specificity was defined as the probability of predicting healthy controls. A positive predictivity value was defined as the probability of HCC if a test result was positive.

Pooled serum samples (n = 6) from HCC patients with high SELDI intensities at 7789 and 7984 m/z were selected. These samples were diluted with 9 M urea, 10 mM Tris/HCI (Ph 7.4) and applied to AKTA Purifier T-900 column system 14. After sample purification, albumin and immunoglobulin were removed from the serum by 3 GA and then by Protein A. The rest of the fractions were loaded onto a Tricine-SDS-PAGE gel according to the methods of Fountoulakis and Schagger 3435, using ETTAN II (Amersham Pharmacia) gel electrophoresis system.

Electrophoresis was run at 20 mA for 3 hr. Gels were then stained with Coomassie Brillian Blue. The bands corresponding to the 8000 m/z markers were excised and then destained with two washes of 50 μl deionised water, followed with 50 μl ACN/50 mM/L NH₄HCO₃ (1:1, v/v), and dried in a SpeedVac concentrator. The dried gel slices were rehydrated with 10 mM DTT followed by 50 mM IAM (45 min at room temperature in the dark). After several washes with 25 mM NH₄HCO₃ and 100% CAN, 20 mg/L-solution of trypsin was added to the gel slices and digestion was allowed to proceed at 37°C for 12 hr.

The trypsin digested sample was loaded onto a C18 reversed-phase column (5 mm × 250 μm, PepMapC18, LC Packings, Amsterdam, The Netherlands), and the peptides were separated by electros pray ionization (ESI, Bruker Esquire 3000, Bruker Daltonik, Bremen, Germany). Proteins were identified by an automated searching algorithm against the SWISS-Protand NCBI protein database.

HCC tissues and adjacent liver tissues (control) were processed according to standard approaches 36. The anti-NAP-2 serum (1:1600, Immunechem Pharmaceuticals INC, Canada) was applied to both HCC and control slides and incubated in a moist chamber at 4°C overnight. 0.01 ml PBS was used as the negative control in all experiments. Slides cut in parallel to the IHC-treated sections were stained by HE for better identification of the different tissue areas. To avoid interindividual bias of IHC staining differentiations, all slides were determined by an experienced pathologist.